<![CDATA[Customer reviews are one of the most valuable sources of feedback for small businesses, but most sentiment analysis systems were not designed to handle regional dialect text. In North Sulawesi, Indonesia, customers of local furniture businesses typically write reviews in Manado Malay dialect (Bahasa Melayu Manado/BMM), using words like "nda" (not), "pe" (very), "so" (already), and "gaga" (good/attractive) that standard Indonesian NLP libraries simply cannot recognize. This research builds and evaluates a web-based sentiment analysis system for CV Talongka jaya, a furniture MSME in Manado, using a Multinomial Naive Bayes Classifier combined with a custom BMM-to-Indonesian lexicon normalization module as the primary contribution. The system was developed using Python with Scikit-learn and CountVectorizer, and Laravel for the web dashboard. A dataset of 250 manually labeled customer reviews was evaluated using both an 80/20 train-test split and 10-Fold Cross Validation. System A (with normalization) achieved 84.00% accuracy, 88.89% precision, 82.76% recall, and 85.71% F1-Score, compared to System B (without normalization) which reached only 66.00% accuracy - an 18-percentage-point improvement. The ablation study confirmed that normalization contributed the single largest performance gain among all preprocessing stages.]]>
Copyrights © 2026
